Machine translation has improved dramatically over the past decade. For many translation tasks — business documents, technical manuals, websites, legal contracts — machine translation with light human review is now genuinely adequate. For books, the picture is more complicated.

This article is an honest assessment of where machine translation helps, where it fails, and what the hybrid approach (machine translation with human post-editing) actually delivers in practice — so you can make an informed decision for your specific book and market.

The Current State of Machine Translation

DeepL is currently the leading machine translation tool for literary prose in European language pairs. Its output is significantly more natural than Google Translate for connected prose — it handles sentence-level context better and produces fewer awkward constructions. ChatGPT and other large language models can also produce surprisingly readable translations and are particularly useful for short passages or for evaluating phrasing options.

For high-resource language pairs (English to French, Spanish, German, Italian, Portuguese, Dutch), DeepL produces output that a skilled post-editor can revise into a readable translation. For lower-resource pairs (English to Finnish, Hungarian, Arabic, many Asian languages), quality drops significantly.

But "readable" and "literary" are different standards. Machine translation consistently produces text that is semantically accurate and grammatically correct while missing almost everything that makes literary writing literary.

Where Machine Translation Works for Books

Technical and Reference Non-Fiction

Non-fiction books with straightforward, informational prose — how-to books, business books, personal finance guides, health and wellness guides — are the strongest candidates for machine-translated + post-edited translation. The prose style is functional rather than artful; the priority is accuracy and clarity rather than voice. A skilled post-editor can produce a publication-ready translation at 40–60% of the cost of a full human translation.

Genre Fiction Where Speed Matters

Some self-published genre fiction authors use machine translation + post-editing for rapid foreign-language release. If you're releasing the fifth book in a romance series to Spanish-language readers who are already fans, a post-edited machine translation may be "good enough" — the readers are invested in the story and characters, and minor translation imperfections won't deter them.

This approach is most defensible for: formulaic genre fiction where prose style is secondary to plot and character, authors who are already established in the target market, and release situations where speed genuinely matters (staying ahead of piracy, maintaining series momentum).

The "good enough" question: What are you comparing to? If the alternative is no translation at all, a post-edited machine translation is clearly better than nothing. If the alternative is a skilled human translator, "good enough" often isn't good enough to build a readership.

Where Machine Translation Fails for Books

Literary Prose

Machine translation handles denotation well and connotation poorly. The difference between two words that mean roughly the same thing — their weight, tone, register, cultural associations — is exactly what makes literary prose literary. Machine translation selects for the statistically most common equivalent rather than the most precise one. The result is prose that is semantically accurate but aesthetically flat.

Literary fiction lives or dies on its prose quality. Readers who pick up a literary novel in Spanish or German and encounter machine-quality prose will not come back for more of your books.

Humor and Wordplay

Machine translation has no reliable mechanism for preserving humor that depends on wordplay, double meanings, or culturally specific references. A pun that works in English will produce a literal translation in French that isn't funny. Dialect humor becomes standard prose. Dry wit becomes flat statement. Machine translation cannot make the creative choices that preserve comedic effect across languages — it can only transfer words.

Cultural References

Contemporary cultural references, idioms, and allusions that are clear to English-speaking readers may be opaque or wrong-footing to readers in other cultures. A skilled human translator makes adaptation decisions: translate the reference, find a cultural equivalent, or leave it as-is with enough context. Machine translation typically either translates literally (losing the reference) or leaves it unchanged (leaving readers confused).

Dialect, Vernacular, and Voice

Character voice differentiation — where one character speaks formally and another in vernacular, where regional dialect carries meaning — is beyond current machine translation. All characters emerge from machine translation sounding approximately the same: neutral, slightly formal, and grammatically standard. This collapses characterization that depends on voice differentiation.

Post-Editing: What It Actually Means

Machine translation post-editing (MTPE) is the process of a human translator revising machine-translated output into a publishable text. It's different from translation — the translator is editing rather than creating — and it's typically faster than translation from scratch.

There are two levels of post-editing:

  • Light post-editing: Correcting errors, removing mistranslations, fixing obvious grammatical problems. The translator doesn't necessarily improve the style — they make the text technically correct. This is appropriate for informational content that doesn't require literary quality. It costs roughly 30–40% of full human translation.
  • Full post-editing: Revising the machine output to produce prose that reads as if it were translated by a skilled human. This includes improving style, adapting cultural references, differentiating character voice, and restructuring awkward sentences. When done well, the result approaches human translation quality. It typically costs 50–70% of full human translation — saving real money on long manuscripts, but less dramatic than authors often expect.

The catch: full post-editing is only as good as the post-editor's literary skills. A translator with strong literary sensibility can take machine output and produce genuinely good prose. A translator who is technically accurate but not a strong literary writer will produce technically accurate but flat prose — similar to the machine output itself. The quality of post-edited translation is almost entirely a function of the human editor's skill.

Specific Tools Worth Knowing

  • DeepL: Best-in-class for European language pairs. Free tier for short texts; subscription for full document translation. The go-to for most literary post-editing workflows.
  • Google Translate: Still widely used, especially for non-European language pairs where DeepL has less coverage. Weaker on literary prose but improving.
  • ChatGPT / Claude: Useful for evaluating translation options, generating alternatives for specific passages, and handling culturally complex segments that puzzle standard MT tools. Not recommended for full-document translation — works better for targeted use on difficult passages.
  • Lara Translate, ModernMT: Adaptive translation tools that can be fine-tuned on your writing style. Niche but potentially valuable for prolific authors translating large volumes of their own work.

The Hybrid Workflow

For authors who want to balance cost and quality, a hybrid workflow can work for the right material:

  1. Run the manuscript through DeepL to generate a first-pass translation.
  2. Hire a skilled literary translator for full post-editing — not just error correction, but genuine prose revision.
  3. Have the post-edited translation read by a native-speaker literary reader who evaluates whether it reads as natural prose in the target language.

This workflow costs roughly 60–70% of a full human translation and, with a skilled post-editor, can produce comparable quality for certain types of prose. For literary fiction with strong voice, wordplay, or dialect — hire a human translator from scratch. The machine output will create as many problems as it solves.

When to Just Hire a Human

Hire a human translator (without machine pre-translation) when:

  • Voice, style, and prose quality are central to the book's value
  • The book contains significant dialect, vernacular, humor, or wordplay
  • You're targeting a literary market where quality signals matter
  • The language pair is poorly supported by current MT tools
  • You're building a long-term readership in a foreign market where your reputation depends on the quality of your translations

The savings from machine translation are real — but they're most real when the quality tradeoff is acceptable for your book and your market. For literary fiction and books where voice is the product, human translation is the only option that protects your reputation.

Before translation — by any method — make sure your source text is as clean and clear as possible. Ambiguous or convoluted sentences in the original become translation problems. EditRoast provides editorial feedback that helps you identify and resolve these issues before they become a translator's headache.